Intelligent detection method for morphology of megakaryocyte of bone marrow

Abstract

The invention discloses an intelligent bone marrow megakaryocyte morphology detection method which comprises the following steps: S1, data preparation: collecting and preprocessing a digital large map of a bone marrow smear, labeling megakaryocytes, and establishing a labeled sample; s2, generating and sorting a sample: extracting a small graph sample by taking megakaryocyte as a center; s3, constructing a deep learning model: constructing and training a deep convolutional neural network, and performing model optimization and performance improvement by using the generated small image sample set; s4, large image detection and reasoning: calling the trained model for reasoning by adopting a sliding window mechanism, and fusing detection results of a plurality of small windows back to an original large image through confidence weighting and a non-maximum suppression strategy; and S5, target cell segmentation: carrying out target segmentation on the detected megakaryocyte, and introducing a pyramid structure for cells with different sizes to obtain an accurate segmentation mask of each target cell. According to the method, the megakaryocyte detection and segmentation precision is improved through large image labeling, small image training and a sliding window reasoning strategy.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a deep learning-based intelligent detection method for bone marrow megakaryocyte morphology, which is applicable to the fields of hematological pathology diagnosis and medical research. Background Art

[0002] Bone marrow megakaryocytes are unique hematopoietic cells in the bone marrow, responsible for platelet production. Measuring 50-100 μm in diameter, they are among the largest hematopoietic cells in the human body. Based on their growth and differentiation, they can be divided into four stages: primitive megakaryocytes (large nuclei with loose chromatin), immature megakaryocytes (with the onset of nuclear lobation), granular megakaryocytes (with granules in the cytoplasm), and platelet-forming megakaryocytes (with platelets forming at the edges of the cytoplasm).

[0003] Under pathological conditions, micromegakaryocytes (diameter <20 μm) or abnormal lobes may appear. Abnormal megakaryocyte numbers or morphology are associated with various diseases. For example, delayed megakaryocyte maturation leading to platelet production disorders can be seen in primary immune thrombocytopenia (ITP), micromegakaryocytes are common in myelodysplastic syndrome (MDS), and significant megakaryocyte proliferation with abnormal morphology can be seen in chronic myeloid leukemia. In myelofibrosis, megakaryocytes are often distributed in clusters and accompanied by reticular fibrosis.

[0004] In addition, low megakaryocyte proliferation can lead to thrombocytopenia, while excessive proliferation is associated with essential thrombocythemia, which is an important indicator for the clinical evaluation of hematopoietic system diseases.

[0005] Currently, clinical hematologists examine bone marrow smears using an optical microscope, first identifying megakaryocytes at low magnification and then determining their classification at high magnification. This completes the megakaryocyte count and differential count across the entire smear, providing evidence for definitive, suggestive, consistent, suspicious, and exclusionary diagnoses. This tedious and time-consuming process requires highly skilled physicians and is subject to significant subjectivity, inefficiency, and poor reproducibility. To improve the quality and efficiency of bone marrow megakaryocyte examinations in clinical diagnosis and treatment, there is an urgent need to enhance the consistency and automation of megakaryocyte counts and differential counts.

[0006] Existing computer-aided detection has the following main problems: megakaryocytes are large in size and difficult to capture in full at one time under high-magnification images; large-image training models are limited by the low proportion of megakaryocytes and significant background interference; conventional image processing requires cutting large images into small images, resulting in the fragmentation of cell structure and loss of morphological features; there is a lack of an effective mechanism for collaborative processing of large and small images, which affects detection accuracy. Summary of the Invention

[0007] To address the above problems, the present invention provides a deep learning-based intelligent detection method for bone marrow megakaryocyte morphology, which significantly improves the cell detection rate and recognition accuracy through the innovative strategies of large-image annotation, small-image training, and rolling window reasoning.

[0008] The technical solution adopted by the present invention to solve its technical problem is: A method for intelligently detecting bone marrow megakaryocyte morphology, comprising the following steps: Step S1. Data preparation stage: By collecting and preprocessing digital large images of bone marrow smears, professional physicians annotate megakaryocytes to create high-quality annotated samples containing location information and morphological categories.

[0009] Furthermore, the digital large images of bone marrow smears are collected and preprocessed, wherein the collected information covers microscope images of patient samples of different ages, genders and disease types; the preprocessing is to evaluate the image clarity by using the Laplace algorithm and reviewed by professional physicians to ensure data quality.

[0010] Step S2. Sample generation and organization: Based on the data prepared in step S1, small image samples are extracted with megakaryocytes as the center, and divided into positive samples, edge samples and negative samples according to their positional relationship.

[0011] Furthermore, the extraction of small image samples is based on the labeled large bone marrow image, and a training sample image block with a size of w×h pixels is generated with the target megakaryocyte as the center.

[0012] Furthermore, the positive sample is a w×h pixel training image block centered on the target megakaryocyte, with typical and complete morphological features and clear and unobstructed images, which is used to train the model to recognize the correct features of standard megakaryocytes.

[0013] Furthermore, the edge samples are image blocks containing some visible megakaryocytes or morphological features at the classification boundary, which are used to enhance the recognition ability and robustness of the model for ambiguous cases and edge situations.

[0014] Furthermore, the negative samples are image blocks of other blood cell types, background tissues or cell fragments that do not contain target megakaryocytes, and are used to train the model to distinguish non-target objects and background interference.

[0015] Step S3. Deep learning model construction: Based on the samples obtained in step S2, a deep convolutional neural network is constructed and trained, including a detection module and a classification module, and the generated small image sample set is used to optimize the model and improve performance.

[0016] Furthermore, the construction and training of the deep convolutional neural network is based on the implementation of a selective layer freezing strategy based on the gradient information of each layer, combined with the AdamW adaptive learning rate optimizer, to effectively improve the stability and generalization performance of the model training.

[0017] Step S4. Large-image detection and reasoning: Based on the model obtained in step S3, a sliding window mechanism with an overlap rate of no less than 20% is used on the new test image. The trained model is called for reasoning, and the detection results of multiple small windows are fused back into the original large image through confidence weighting and non-maximum suppression strategies.

[0018] Furthermore, the sliding window mechanism with an overlap rate of not less than 20% is used to prevent the integrity of cells from being destroyed when a large image is cut into small images, avoid the loss of feature information, and ensure the accuracy of judgment.

[0019] Step S5. Target cell segmentation: Based on the inference results of step S4, the detected megakaryocytes are segmented. A pyramid structure is introduced for cells of different sizes to obtain an accurate segmentation mask for each target cell, thereby realizing intelligent morphological detection of bone marrow megakaryocytes.

[0020] Furthermore, the precise segmentation mask is to refine the mask edge of the segmentation result, and use edge detection and smoothing technology to optimize the contour continuity and accuracy.

[0021] Furthermore, the pyramid structure strategy is a multi-scale feature fusion technology that addresses the problem of target size variation by constructing feature pyramids at different resolution levels, thereby significantly improving the segmentation accuracy of small cells and adherent cells.

[0022] The beneficial effects of the present invention are as follows: the present invention improves the accuracy of megakaryocyte detection and segmentation through large-image annotation, small-image training and sliding window reasoning strategies; collects diverse high-quality data and combines it with professional review to enhance training reliability; introduces deep learning models and optimization strategies to improve model stability and generalization capabilities; and realizes automated counting to assist in the diagnosis of blood diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is an overall flow chart for implementing the present invention.

[0024] Figure 2 Schematic diagram of microscope camera image acquisition.

[0025] Figure 3 This is a labeled schematic diagram of megakaryocytes in the large picture.

[0026] Figure 4 This is a schematic diagram of cutting a small image on a large image based on the location of megakaryocytes.

[0027] Figure 5 It is a schematic diagram of the convolutional neural network (CNN) model structure.

[0028] Figure 6 This is a schematic diagram of cutting small graphs with a rolling window for reasoning.

[0029] Figure 7 It is a schematic diagram of the inference results. DETAILED DESCRIPTION

[0030] The following is combined with Figure 1-6 And specific implementation method further details: like Figure 1 As shown in the figure, the flowchart describes the complete process of intelligent detection of bone marrow megakaryocyte morphology based on deep learning, which includes five main steps (S1-S5): S1. Data preparation stage Step 1.1: Data collection and organization 1.1.1 Image source acquisition: Bone marrow smear microscopy images were collected from patients of different age groups, genders, and disease types; the diversity and representativeness of the sample sources were ensured to avoid data bias.

[0031] Step 1.2: Image quality screening 1.2.1 Automated quality assessment: An image clarity detection algorithm is used to calculate the Laplace variance to assess the degree of blur. A clarity threshold is then set to automatically mark suspected blurry images. A brightness distribution histogram is also created to identify overexposed and underexposed images, thereby completing image quality screening.

[0032] 1.2.2 Manual quality review: Experienced professional physicians will manually review the automatic screening results; then a unified quality assessment standard and scoring system will be established.

[0033] 1.2.3 Abnormal sample processing: Establish an image quality abnormality classification system covering blur, overexposure, underexposure, uneven coloring, etc., and completely eliminate low-quality images that cannot be repaired.

[0034] Step 1.3: Standardize the preprocessing process 1.3.1 Consistency Verification: Check the consistency of pre-processed images through random sampling, calculate the distribution of basic parameters such as image brightness, contrast, and color saturation, ensure the standardization effect of images from different batches and sources, and establish quality control checkpoints and acceptance criteria.

[0035] S2. Generate and organize samples obtained in S1 Step 2.1: Annotated data parsing and preprocessing 2.1.1 Annotation Information Extraction: Parse the JSON format file annotated by professional physicians to obtain the precise coordinate information of megakaryocytes, extract additional information such as the bounding box coordinates (w_min, h_min, w_max, h_max) and category label of each cell, and verify the integrity and format consistency of the annotated data.

[0036] 2.1.2 Coordinate system standardization: Unify the coordinate systems of different annotation tools to the upper left corner origin, convert relative coordinates into absolute pixel coordinates, handle the coordinate scaling problem of images with different resolutions, and establish a coordinate verification mechanism to ensure the accuracy of the annotation position.

[0037] Step 2.2: Megakaryocyte-centered sample extraction 2.2.1 Center point calculation and positioning: The geometric center point coordinates of megakaryocytes are calculated based on the annotated bounding box: center_w = (w_min + w_max) / 2 and center_h = (h_min + h_max) / 2. Considering the irregularity of cell shape, the center point positioning is optimized using the center of mass calculation method, and a center point offset tolerance mechanism is established.

[0038] 2.2.2 Generate a w × h pixel image block based on the calculated center point. Expand the image block by h / 2 and w / 2 pixels in the upper, lower, left, and right directions respectively. Determine the extraction boundaries as block_w_min = center_w - w / 2, block_w_max = center_w + w / 2, block_h_min = center_h - h / 2, and block_h_max = center_h + h / 2. Use the difference displacement strategy to handle boundary violations to ensure that the size of all extracted image blocks is strictly w × h pixels.

[0039] Step 2.3: Sample quality control and verification 2.3.1 Automated quality detection, including image quality assessment to detect the clarity, contrast, and color saturation of extracted samples, content integrity verification to confirm the integrity and identifiability of cell structures, duplicate sample detection and removal using perceptual hashing technology, and abnormal sample tagging to automatically mark samples with abnormal quality or content.

[0040] 2.3.2 Expert review and annotation verification, including professional physicians conducting random sampling quality review of 10%-20% of samples, verifying the consistency of sample category labels with the actual content, discussing the boundary conditions of fuzzy classification samples, and unifying the review standards through standardized training to ensure the consistency of different experts' judgments. After the above steps, the data is processed to obtain a standardized data set.

[0041] S3. Deep learning model construction Step 3.1: Network architecture design and construction 3.1.1 Backbone network selection and configuration, including architecture selection analysis comparing the performance of mainstream networks such as ResNet and RegNet, depth configuration optimization to select an appropriate network depth (usually 50-101 layers) based on the complexity of megakaryocyte features, channel number adjustment to optimize the channel configuration of each layer based on the cell image feature dimension, and loss function selection using an adaptive FocalLoss loss function to address the class imbalance problem.

[0042] Step 3.2: Gradient information analysis and layer freezing strategy 3.2.1 The gradient information monitoring system includes real-time calculation of gradient statistics of the mean, variance, and norm of the weight gradient of each layer, gradient flow visualization that plots the gradient propagation in the network and identifies layers with vanishing or exploding gradients, layer importance assessment based on the gradient amplitude to evaluate the contribution of each layer to the final loss, and a dynamic monitoring mechanism for gradient change trend analysis during training.

[0043] 3.2.2 Selective layer freezing decision includes setting the freezing threshold to consider freezing when the gradient norm is <1e-6, adopting a layer freezing strategy that freezes shallow feature extraction layers with smaller gradient updates and keeps deep features closely related to megakaryocyte detection trainable, and a dynamic freezing adjustment mechanism that dynamically adjusts the freezing strategy according to the training progress.

[0044] Step 3.3: AdamW optimizer configuration and tuning 3.3.1 Adaptive learning rate scheduling, including learning rate warmup that uses a small learning rate at the beginning of training and gradually increases it to a set value, a cosine annealing strategy using CosineAnnealingLR to achieve smooth learning rate decay, platform learning rate scheduling that automatically reduces the learning rate when the validation loss stops decreasing, and an early stopping mechanism that prevents overfitting and stops training when the validation performance no longer improves.

[0045] Step 3.4: Training strategy optimization and performance improvement 3.4.1 Loss function design, including detection loss that combines classification loss, bounding box regression loss, and confidence loss, classification loss that uses FocalLoss to handle class imbalance, multi-task learning that balances the loss weights of detection and classification tasks, and loss weight adjustment that dynamically adjusts the weights of each loss based on task importance.

[0046] 3.4.3 Regularization techniques, including detection loss that combines classification loss, bounding box regression loss, and confidence loss, using FocalLoss to handle classification loss with class imbalance, multi-task learning that balances the loss weights of detection and classification tasks, and a loss weight adjustment mechanism that dynamically adjusts the weights of each loss based on task importance.

[0047] Step 3.5: Training Monitoring and Evaluation 3.5.1 Training process monitoring, including loss function tracking to monitor the changing trends of training loss and validation loss in real time, accuracy monitoring to track classification accuracy and detection accuracy, recording learning rate adjustment history and learning rate changes of optimizer status, and gradient statistics to monitor gradient norm changes and promptly detect training anomalies.

[0048] S4. Large-scale image detection and reasoning methods Step 4.1: Design and implementation of sliding window mechanism for small graph cutting 4.1.1 Small Image Window Parameter Configuration: In this embodiment of the present invention, a sliding window method is used to process large images and perform small image segmentation, which specifically includes the following parameter configuration steps: Window size setting: Determine the width and height of the sliding window based on the input size of the training model used to ensure the consistency of the input image; Overlap rate calculation: Set the overlap rate between windows to 20% to 30%, and calculate the sliding step size of adjacent windows to ensure that the cell edge area can be fully detected; Adaptive parameter adjustment: Dynamically adjust the window size and sliding step size based on the input image size and cell density to adapt to detection requirements in different image scenarios.

[0049] Step 4.2: Confidence-weighted fusion strategy for small-graph inference results 4.2.1 Overlapping Region Identification and Grouping: To avoid duplicate detection caused by the sliding window, the present invention identifies overlapping regions and groups them together on the small graph inference results. This specifically includes: Overlap detection algorithm: Use the overlap judgment algorithm to calculate the degree of spatial overlap between detection frames in different windows; Overlap threshold setting: Set the intersection-over-union (IoU) threshold to determine whether multiple detection boxes point to the same target cell; Clustering and grouping processing: multiple detection frames that meet the overlapping conditions are grouped into the same group, and each group corresponds to a real cell; Spatial distribution analysis: Combine the spatial location relationship of the detection results to further improve the accuracy of duplicate detection and identification.

[0050] 4.2.2 Confidence Weighted Fusion Algorithm Perform weighted fusion processing on the identified duplicate detection results, including: Confidence weight calculation: Calculate the fusion weight based on the confidence score of each detection box to increase the contribution of high-confidence results; Coordinate weighted averaging: Perform weighted averaging on the coordinate information of the detection boxes within the cluster group to generate the coordinates of the fused bounding box; Confidence update mechanism: The confidence of the fusion result is updated according to the weighted logic to reflect its comprehensive reliability; Category consistency processing: When the categories of multiple detection results of the same cell are inconsistent, the majority voting or highest confidence judgment strategy is used to unify their final category labels.

[0051] S5. Target Cell Segmentation 5.1 Instance Segmentation and Boundary Refinement: Cell contours are extracted using a deep learning-based instance segmentation model to obtain accurate segmentation masks for each target cell: The instance segmentation model (Instance Segmentation): uses the mainstream instance segmentation network of Mask R-CNN, combined with the backbone network ResNet to extract high-level semantic features to achieve independent segmentation of each cell instance.

[0052] Mask Refinement: Refines the mask edges in the segmentation results, including using edge detection and edge smoothing to optimize contour continuity and accuracy.

[0053] Multi-scale fusion (Optional): To cope with cells of different sizes, a pyramid structure strategy is adopted to improve the segmentation accuracy of small cells or adherent cells.

[0054] like Figure 2 As shown, the hardware device used in the autofocus method is a microscope system equipped with an autofocus function, which includes the following key components: Camera (1): installed on the top of the microscope, used to collect images in the microscope field of view and transmit them to a computer or other processing equipment for analysis. Objective lens (2): installed on the objective lens turntable, responsible for magnifying the sample. It usually contains multiple objective lenses with adjustable magnification, and users can switch between different magnifications according to their needs. Motorized stage (3): used to place slides and capable of high-precision XY axis movement to achieve automatic scanning and alignment of samples. High-power microscope light source (4): provides strong illumination to ensure sufficient image brightness during the focusing process, reduce noise, and improve the accuracy and stability of autofocus. Automatic Z axis (5): controls the vertical movement of the objective lens or stage to accurately adjust the focal length and achieve autofocus function.

[0055] The system combines camera image acquisition, automatic Z-axis focal length adjustment, and deep learning models for focus optimization to achieve high-precision autofocus function, which is suitable for application scenarios such as microscope scanning and pathological analysis.

[0056] like Figure 3The image shown above shows a bone marrow smear sample collected under a microscope, displaying a large number of cells of varying morphologies densely packed within the field of view, reflecting the complex microstructure of bone marrow tissue. The sample contains a variety of hematopoietic cells, among which megakaryocytes stand out due to their large size (up to 50-100μm in diameter) and unique morphological features. A specific area in the image has been clearly selected by a professional physician using a rectangular frame or a specific marker, clearly indicating the location of the megakaryocyte. This selected area typically highlights typical megakaryocyte features, such as large cell size, multilobed nuclear structure, or richly granular cytoplasm, providing an accurate reference for subsequent morphological classification and analysis. This labeling process, performed by a professional hematologist, ensures the accuracy and reliability of the labeling, providing a high-quality data foundation for training deep learning-based megakaryocyte detection and classification models.

[0057] like Figure 4 The image shows a bone marrow smear sample collected using a high-resolution microscope. It demonstrates the dense distribution of cells within the bone marrow tissue, encompassing a wide range of hematopoietic cell types and highlighting the complexity of the bone marrow microenvironment. The image accurately annotates the center of a megakaryocyte with white rectangular boxes. These boxes were selected by hematologists to highlight typical megakaryocyte features, such as large cell size (50-100μm in diameter), multilobed nuclear structure, and richly granular cytoplasm, ensuring accurate and reliable annotation. Surrounding these white boxes, the image is further marked with black rectangular boxes, indicating a small area cropped from the megakaryocyte center. These black boxes are typically generated with a fixed size (e.g., w × h pixels) to extract a small sample containing the complete megakaryocyte structure while preserving the necessary surrounding background information to support subsequent deep learning model training and analysis. This cropping strategy effectively avoids information loss caused by low cell density or background noise when directly processing the large image, providing high-quality training data for accurate megakaryocyte detection and classification.

[0058] like Figure 5 Figure 2 shows a convolutional neural network (CNN) model architecture for classification tasks, which is designed for the instance segmentation task of cells in images.

[0059] Input layer: The input layer receives an image, which is a tissue slice. This image may be an image of cells or other structures. Different colors (purple and pink) may represent different cell types or tissues. The image is converted into a two-dimensional matrix of pixel values.

[0060] Convolutional layer (feature map): The input image passes through the convolutional layer, which extracts features from the image. The convolutional layer slides multiple convolution kernels (filters) over the image, performing mathematical operations (convolution) to detect features such as edges, corners, and textures in the image.

[0061] Each convolution kernel extracts different features, and after convolution, multiple feature maps are generated. These feature maps represent the performance of the image in different feature dimensions.

[0062] Pooling Layer: After the convolutional layers, the model applies a pooling layer, typically a max pooling operation. The purpose of pooling is to reduce the spatial dimensionality of the feature map, lowering computational complexity and mitigating the risk of overfitting. Max pooling selects the maximum value within each region it slides over, retaining the most important features. Pooling helps reduce model complexity while preserving key features in the image.

[0063] Flattening layer: After the pooling layer, the feature map is flattened into a one-dimensional vector. This is to convert the data into a form that can be processed by the fully connected layer, because the fully connected layer requires one-dimensional input data.

[0064] Fully connected layer: A fully connected layer connects the flattened output to a layer of neurons. Each neuron is connected to all neurons in the previous layer. The purpose of a fully connected layer is to learn the complex relationships between the features extracted by the convolutional layer and further understand the patterns in the image.

[0065] Output layer: The last layer is the output layer, where the model makes the final prediction. The cell classification in the output image is Figure 5 There are multiple output nodes (labeled 0, 1, 2, 3, etc.). During training, AdaptiveFocalLoss is used as the loss function (see below), and the model parameters are optimized through backpropagation and gradient descent algorithms.

[0066] like Figure 6 The image shown in Figure 1 shows a bone marrow smear sample under a microscope, which is used for cell recognition tasks. Considering that the model training phase uses a fixed-size input image, the inference phase uses a sliding window method to divide the entire large image into blocks.

[0067] To ensure feature continuity and coverage integrity, the sliding window is set as follows: Window size: Consistent with the model input size during training, w × h. Overlap: Set to 20%–30% to enhance the model's ability to perceive edge regions and reduce edge prediction errors. Stride calculation: Horizontal stride (stride_w) = w × (1-overlap), vertical stride (stride_h) = h × (1-overlap).

[0068] Figure 6 The diagram in the upper left corner shows how the sliding windows are cropped and their overlapping areas.

[0069] During inference, the entire image is cropped into several overlapping smaller images, which are fed into the model sequentially for prediction. Finally, the prediction results of all the smaller images are concatenated or fused into the inference result for the entire image.

[0070] like Figure 7 The image shows a schematic diagram of the inference results of intelligent bone marrow megakaryocyte morphology detection. The image contains eight sub-images, divided into two rows of four sub-images each, demonstrating the deep learning model's ability to classify megakaryocyte morphology. Each sub-image is labeled below the megakaryocyte class: "1" indicates a primitive megakaryocyte (large nucleus and loose chromatin), "2" indicates an immature megakaryocyte (nuclear lobation is beginning), "3" indicates a granular megakaryocyte (cytoplasm filled with granules), and "4" indicates a platelet-producing megakaryocyte (platelets are formed at the rim of the cytoplasm). The four sub-images in the first row, all labeled "2," demonstrate typical features of immature megakaryocytes, such as medium size and incipient nuclear lobation. The first sub-image in the second row, labeled "1," shows the large nucleus and loose chromatin characteristic of primitive megakaryocytes. The last three sub-images in the second row, labeled "3," highlight the granular cytoplasm of granular megakaryocytes. These sub-images are small samples of bone marrow smears under a microscope, generated using a rolling window inference strategy. They demonstrate the model's ability to accurately classify megakaryocyte morphology on different samples, providing reliable visual support for blood pathology diagnosis.

[0071] Technical advantages and application results of this invention: 1. Improved detection efficiency and accuracy: Through innovative strategies of large-image annotation, small-image training, and rolling window inference, the detection rate and morphological recognition accuracy of bone marrow megakaryocytes are significantly improved, overcoming the limitations of traditional methods such as strong subjectivity, low efficiency, and poor repeatability.

[0072] 2. Enhanced data quality: Large digital images of bone marrow smears covering different ages, genders, and disease types are collected. The image clarity is evaluated using the Laplace algorithm and reviewed by professional physicians to ensure high-quality labeled samples and improve the reliability of model training.

[0073] 3. Optimize model performance: Utilize deep convolutional neural networks, combined with a selective layer freezing strategy and the AdamW optimizer, to improve the stability and generalization capabilities of model training and adapt to complex pathological scenarios.

[0074] 4. Solve the problem of large-image processing: A sliding window mechanism with an overlap rate of no less than 20% is used to avoid cell structure fragmentation and feature loss caused by large-image cutting, ensuring the integrity and accuracy of detection results.

[0075] 5. Improve segmentation accuracy: Using a deep learning-based instance segmentation model, combined with edge detection, smoothing technology, and pyramid structure or multi-scale context fusion strategy, it significantly improves the segmentation accuracy of small cells and adhesion cells, and optimizes contour continuity and accuracy.

[0076] 6. Promote clinical application: Achieve automation and consistency in megakaryocyte counting and differential counting, reduce physician workload, provide efficient and reliable auxiliary tools for hematopathology diagnosis and medical research, and facilitate accurate diagnosis of diseases.

Claims

1. A method for intelligent detection of bone marrow megakaryocyte morphology, characterized by: The following steps are included Step S1. Data preparation: By collecting and preprocessing digital images of bone marrow smears, professional physicians annotate megakaryocytes to create high-quality annotated samples containing location information and morphological categories; Step S2. Sample generation and organization: Based on the data prepared in step S1, small image samples are extracted with megakaryocytes as the center, and divided into positive samples, edge samples, and negative samples according to their positional relationship; Step S3. Deep Learning Model Construction: Based on the samples obtained in step S2, a deep convolutional neural network is constructed and trained, including a detection module and a classification module. The generated small image sample set is used to optimize the model and improve performance. Step S4. Large-image detection and inference: Based on the model obtained in step S3, a sliding window mechanism with an overlap rate of at least 20% is used on the new test image. The trained model is called for inference, and the detection results of multiple small windows are fused back into the original large image through confidence weighting and non-maximum suppression strategies. Step S5. Target cell segmentation: Based on the inference results of step S4, the detected megakaryocytes are segmented. A pyramid structure is introduced for cells of different sizes to obtain an accurate segmentation mask for each target cell, thereby realizing intelligent morphological detection of bone marrow megakaryocytes.

2. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S1, a digital large image of a bone marrow smear is collected and pre-processed, wherein the collected information includes microscope images of patient samples of different ages, genders and disease types; The preprocessing is to evaluate the image clarity by using the Laplace algorithm and to be reviewed by professional physicians to ensure data quality.

3. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S2, the extraction of small image samples is based on the labeled large bone marrow image, and a training sample image block with a size of w×h pixels is generated with the target megakaryocyte as the center.

4. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S2, the positive sample is a w×h pixel training image block centered on the target megakaryocyte, with typical and complete morphological features and clear and unobstructed images, which is used to train the model to recognize the correct features of standard megakaryocytes.

5. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S2, the edge samples are image blocks containing some visible megakaryocytes or morphological features at the classification boundary, which are used to enhance the recognition ability and robustness of the model for ambiguous cases and edge cases.

6. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S2, the negative samples are image blocks of other blood cell types, background tissues or cell debris that do not contain target megakaryocytes, and are used to train the model to distinguish non-target objects and background interference.

7. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S3, the construction and training of the deep convolutional neural network is based on the implementation of a selective layer freezing strategy based on the gradient information of each layer, combined with the AdamW adaptive learning rate optimizer, to effectively improve the stability and generalization performance of the model training.

8. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S4, the sliding window mechanism with an overlap rate of not less than 20% is adopted to prevent the integrity of cells from being destroyed when the large image is cut into small images, avoid the loss of feature information, and ensure the accuracy of judgment.

9. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S5, the precise segmentation mask is to refine the mask edge of the segmentation result, and use edge detection and smoothing technology to optimize the contour continuity and accuracy.

10. The method for intelligent detection of bone marrow megakaryocyte morphology according to claim 1, characterized in that: In step S5, the pyramid structure strategy is a multi-scale feature fusion technology that handles the problem of target size variation by constructing feature pyramids at different resolution levels, thereby significantly improving the segmentation accuracy of small cells and adherent cells.

Images